Reproduce “TimeMar: Multi-Scale Autoregressive Modeling for Unconditional Time Series Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiangyu Xu, Qingsong Zhong, Jilin Hu
TimeMAR is a generative framework for unconditional time series generation that explicitly models multi-scale temporal patterns and structural heterogeneity in time series data. The approach achieves superior generation quality, reducing the Discriminative Score by up to 50% on several datasets while demonstrating high parameter efficiency and robust performance for long sequences.
Key insights
- Decomposing time series into distinct trend and seasonal components allows for specialized modeling, enhancing the generation of heterogeneous temporal patterns.
- Multi-scale tokenization via a VQ-VAE enables coarse-to-fine generation, effectively capturing hierarchical temporal dependencies in the data.
- Utilizing frequency-domain features and coarse-guided reconstruction improves the stability and fidelity of generating challenging high-frequency, non-stationary seasonal components.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.